Estimating Mode Choice in Decentralized Shared Mobility: A Bagging-Enhanced Heterogeneous Ensemble Method

Key Takeaways

• BESHEM combines linear, tree, probabilistic, instance, neural models
• Outperforms 20 base models and four ensemble benchmarks
• Extra Trees meta-learner yields highest accuracy
• Adoption driven by prior ride‑sharing, metro trips, female safety
• Privacy concerns and short trips deter UPR usage

Pulse Analysis

Ensemble learning has become a cornerstone of predictive analytics, yet most transportation studies rely on single‑model approaches that struggle with the high nonlinearity of traveler behavior. By nesting bagging within a stacking framework, BESHEM leverages the variance reduction of bagging and the bias correction of stacking, creating a heterogeneous architecture that adapts to diverse data patterns. This methodological leap addresses the long‑standing trade‑off between model interpretability and predictive power, offering a scalable template for other complex choice contexts such as multimodal trip planning or dynamic pricing.

The authors test BESHEM on a rich dataset of User‑organized Pre‑pooled Ride‑hailing (UPR) participants across several suburban university campuses in China. The data capture socioeconomic profiles, travel scenarios, and attitudinal variables, allowing the model to uncover nuanced drivers of adoption. Results show that prior exposure to ride‑sharing, longer metro‑linked trips, and heightened safety perception among women significantly boost UPR uptake, while concerns over privacy and the availability of short‑distance alternatives dampen interest. The Extra Trees meta‑learner, with its ability to handle high‑dimensional feature interactions, emerges as the most effective aggregator, pushing prediction accuracy beyond that of any individual algorithm.

For industry stakeholders, the implications are twofold. First, the superior forecasting accuracy equips mobility providers with actionable insights to tailor marketing, pricing, and service design to the most receptive user segments. Second, urban planners can integrate these refined demand estimates into multimodal network simulations, ensuring that infrastructure investments align with emerging shared‑mobility trends. As AI‑driven ensembles like BESHEM gain traction, they promise to bridge the gap between academic rigor and real‑world decision‑making, fostering smarter, more resilient transportation ecosystems.